muzzy/GLM-5.3-GGUF

This repo contains the same cooked quants that Ubergarm made for GLM 5.1, just for GLM 5.3 now.

Status 09/25: uploading quants, perplexity coming soon after.

Imatrix calculated with:

./build/bin/llama-imatrix -m ~/GLM-5.3/GLM-5.3-00001-of-00115.gguf -f ~/ubergarm-imatrix-calibration-corpus-v02.txt -o ~/ubergarm-imatrix.dat --defer-experts -ngl 999 --layer-similarity --ctx-size 512

Perplexity (wikitext-103 test split, 289,280 tokens / 565 chunks, ctx 512, ubatch 512, flash attention on, CPU-only 112 threads):

Quant	PPL	+/-
IQ1_KT	4.7472	0.02803
IQ2_KS	3.9404	0.02232
IQ2_KL	in progress	-
IQ3_KS	-	-
IQ4_K	-	-
Q4_0_R8	-	-
Q4_K_R4	-	-
Q6_K_R4	-	-
Q8_K_R8	-	-
Q8_0_R8	-	-

Calculated with:

./build/bin/llama-perplexity -m <QUANT>/GLM-5.3-<QUANT>-00001-of-00015.gguf -f wiki.test.txt --ctx-size 512 --ubatch-size 512 -fa on --seed 1337 --threads 112